Unogona here kushandisa AI kugadzira rekodhi nyowani kubva kune waunofarira muimbi?
Zvichangoburwa mukudzidza kwemichina zvakaratidza kuti modhi dzave kukwanisa kunzwisisa data yakaoma senge zvinyorwa nemifananidzo. OpenAI's Jukebox inoratidza kuti kunyangwe mimhanzi inogona kuenzanisirwa chaizvo neneural network.
Mimhanzi chinhu chakaoma kuenzanisira. Iwe unofanirwa kufunga nezve ese ari maviri maficha akadai se tempo, ruzha, uye kurira uye zvimwe zvakaomarara zvakadai semazwi, zviridzwa, uye mimhanzi chimiro.
Kushandisa advanced machine learning matekiniki, OpenAI yawana nzira yekushandura mbishi audio kuita inomiririra inogona kushandiswa nemamwe mamodheru.
Ichi chinyorwa chinotsanangura izvo Jukebox inogona kuita, mashandiro ayo, uye zvipimo zvazvino zvetekinoroji.
Chii chinonzi Jukebox AI?
Jukebox ndeye neural net modhi neOpenAI iyo inogona kugadzira mimhanzi nekuimba. Iyo modhi inogona kugadzira mimhanzi mumhando dzakasiyana siyana uye maartist masitaera.
Semuenzaniso, Jukebox inogona kuburitsa rwiyo rwedombo muchimiro chaElvis Presley kana hip hop tune mumhando yeKanye West. Unogona kushanyira izvi Website kuongorora kuti iyo modhi inoshanda sei pakubata ruzha rwevaunofarira mimhanzi maartist nemhando.
Iyo modhi inoda genre, muimbi, uye mazwi sekuisa. Iyi inopinza inotungamira modhi yakadzidziswa pamamirioni evatambi uye data renziyo.
Jukebox inoshanda sei?
Ngatitarisei kuti Jukebox inokwanisa sei kugadzira riwaya redhiyo kubva kumhando yakadzidziswa pamamirioni enziyo.
Encoding process
Nepo mamwe mamodheru erudzi rwemimhanzi anoshandisa MIDI yekudzidziswa data, Jukebox inodzidziswa pane chaiyo mbishi faira redhiyo. Kumanikidza odhiyo munzvimbo yakasarudzika, Jukebox inoshandisa otomatiki encoder nzira inozivikanwa seVQ-VAE.
VQ-VAE inomirira Vector Quantized Variational Autoencoder, iyo inogona kunzwika seyakaoma, saka ngatizviparadze.
Kutanga, ngatiedzei kunzwisisa zvatinoda kuita pano. Kuenzaniswa nemazwi kana mimhanzi yepepa, faira reodhiyo rakaomeswa rakanyanya kuomarara. Kana tichida kuti modhi yedu "idzidze" kubva munziyo, tichafanirwa kuishandura kuita inomiririra yakamanikidzwa uye yakapfava. In machine learning, tinodaidza chinomiririra ichi a nzvimbo yakanyarara.
An autoencode inzira yekudzidza isina anotariswa inoshandisa a neural network kutsvaga isiri-mutsara yakamira inomiririra yeakapihwa data kugovera. Iyo autoencoder ine zvikamu zviviri: encoder uye decoder.
The encoder inoedza kutsvaga nzvimbo yakadzikama kubva kune imwe seti yedata raw apo iyo Decoder inoshandisa iyo yakamira inomiririra kuedza kuigadzira patsva kuti idzokere muchimiro chayo chepakutanga. Iyo autoencoder inodzidza nzira yekudzvanya iyo data mbishi nenzira inoderedza kukanganisa kwekuvakazve.
Zvatave kuziva zvinoitwa neautoencoder, ngatiedzei kunzwisisa zvatinoreva ne "variational" autoencoder. Kuenzaniswa neyakajairwa autoencoder, akasiyana autoencoders anowedzera isati yasvika nzvimbo yakadzikama.
Pasina kunyura mumasvomhu, kuwedzera probabilistic pamberi inochengeta kugovaniswa kwakadzikama kwakabatana. Musiyano mukuru pakati peVAE neVQ-VAE ndewekuti iyo yekupedzisira inoshandisa discrete latent inomiririra pane inoenderera.
Yega yega VQ-VAE nhanho inozvimiririra inoisa iyo yekuisa. Iyo yepasi nhanho encoding inoburitsa yakanyanya-mhando yekuvakazve. Iyo yepamusoro-level encoding inochengeta yakakosha ruzivo rwemimhanzi.
Kushandisa Transformers
Iye zvino zvatine mimhanzi makodhi encoded neVQ-VAE, tinogona kuedza gadzira mimhanzi munzvimbo ino yakadzvanywa discrete.
Jukebox inoshandisa autoregressive transformers kugadzira odhiyo inobuda. Transformers imhando yeneural network inoshanda zvakanyanya neyakateedzana data. Zvichipa kutevedzana kwezviratidzo, mushanduri weshanduro anoedza kufanotaura chiratidzo chinotevera.
Jukebox inoshandisa yakareruka musiyano weSparse Transformers. Kana mamodheru ese ekutanga adzidziswa, shanduri inogadzira makodhi akamanikidzwa ayo anobva adzoserwa kumashure muodhiyo yakapfava uchishandisa VQ-VAE decoder.
Artist uye Genre Conditioning muJukebox
Jukebox's generative modhi inoitwa zvakanyanya kudzoreka nekupa mamwe masaini masaini panguva yekudzidziswa nhanho.
Mamodheru ekutanga anopihwa nevatambi uye mazita emhando yerwiyo rumwe norumwe. Izvi zvinoderedza entropy yekufanotaura kweodhiyo uye inobvumira modhi kuti iwane mhando iri nani. Iwo mavara anotigonesawo kutungamira modhi mune imwe nzira.
Kunze kwemuimbi uye genre, zvikwangwani zvenguva zvinowedzerwa panguva yekudzidziswa. Aya masaini anosanganisira kureba kwerwiyo, nguva yekutanga yemuenzaniso, uye chikamu cherwiyo rwapfuura. Iyi yekuwedzera ruzivo inobatsira modhi kuti inzwisise maodhiyo mapatani anovimba neiyo yakazara chimiro.
Semuenzaniso, modhi inogona kudzidza kuti kuombera kwemimhanzi mhenyu kunoitika panopera rwiyo. Iyo modhi inogona zvakare kudzidza, semuenzaniso, kuti mamwe genre ane zvikamu zvezviridzwa zvakareba kupfuura mamwe.
Lyrics
Iwo maconditioned models akataurwa muchikamu chakapfuura anokwanisa kugadzira akasiyana manzwi ekuimba. Zvisinei, manzwi aya anowanzova asina kubatana uye asingazivikanwe.
Kudzora iyo generative modhi kana zvasvika kune lyric chizvarwa, vaongorori vanopa mamwe mamiriro panguva yekudzidziswa. Kubatsira mepu lyric data kunguva pane chaiyo odhiyo, vaongorori vakashandisa Spleeter kubvisa mazwi uye NUS AutoLyricsAlign kuti uwane kurongeka kwemazwi emazwi.
Kuganhurirwa kweJukebox Model
Imwe yemipimo mikuru yeJukebox ndiko kunzwisisa kwayo kwezvimiro zvakakura zvemimhanzi. Semuyenzaniso, kapfupi kapfupi 20-second clip yezvinobuda inganzwika kunakidza, asi vateereri vanozoona kuti chimiro chemimhanzi chekudzokorora makorasi nemavhesi hachipo mukubuda kwekupedzisira.
Iyo modhi zvakare inononoka kupa. Zvinotora anenge maawa 9 kupa zvizere miniti imwe yekuteerera. Izvi zvinodzikamisa nhamba yenziyo dzinogona kugadzirwa uye inodzivirira iyo modhi kuti isashandiswe mukupindirana maapplication.
Chekupedzisira, vaongorori vakacherekedza kuti dhatabheti remuenzaniso rinonyanya muChirungu uye rinoratidza kunyanya magungano emimhanzi yekuMadokero. Vatsvagiri veAI vanogona kutarisa kutsvaga kweramangwana pakugadzira mimhanzi mune mimwe mitauro uye isiri yekuMadokero mimhanzi masitaera.
mhedziso
Iyo Jukebox purojekiti inosimbisa kugona kuri kukura kwemamodhi ekudzidza emuchina kugadzira yakaringana inomiririra yedata rakaomarara senge mbishi audio. Kubudirira kwakafanana kuri kuitika muzvinyorwa, sezvinoonekwa mumapurojekiti akadai GPT-3, uye mifananidzo, sezvinoonekwa muOpenAI's DALL-E2.
Kunyange tsvakiridzo munzvimbo ino yave inonakidza, pachine kushushikana pamusoro pekodzero dzepfuma yehungwaru uye mabatiro angave nemamodheru aya pamaindasitiri ekugadzira zvakazara. Vatsvagiri nevagadziri vanofanirwa kuenderera mberi nekubatana kuitira kuti mamodheru aya arambe achivandudza.
Mamodheru emumhanzi anogadzirwa anogona munguva pfupi kuita sechishandiso chevaimbi kana sechishandiso chevagadziri vanoda mimhanzi yakasarudzika yemapurojekiti.
Leave a Reply